使用 OpenMP 进行归约以计算矩阵元素的最终总和值

Question

I have the following double loop where I compute the element of matrix Fisher_M[FX][FY] . I tried to optimize it by putting a OMP pragma #pragma omp parallel for schedule(dynamic, num_threads) but gain is not as good as expected.

Is there a way to do a reduction witht OpenMP (of sum) to compute quickly the element Fisher_M[FX][FY] ? Or maybe this is doable with MAGMA or CUDA ?

#define num_threads 8

#pragma omp parallel for schedule(dynamic, num_threads)
for(int i=0; i<CO_CL_WL.size(); i++){
    for(int j=0; j<CO_CL_WL.size(); j++){
        if( CO_CL_WL[i][j] != 0 || CO_CL_WL_D[i][j] != 0){
          Fisher_M[FX][FY] += CO_CL_WL[i][j]*CO_CL_WL_D[i][j];
        }
    }
}

Answer 1

Your code has a race condition at line Fisher_M[FX][FY] += ... . Reduction can be used to solve it:

double sum=0;  //change the type as needed
#pragma omp parallel for reduction(+:sum) 
for(int i=0; i<CO_CL_WL.size(); i++){
    for(int j=0; j<CO_CL_WL.size(); j++){
        if( CO_CL_WL[i][j] != 0 || CO_CL_WL_D[i][j] != 0){
          sum += CO_CL_WL[i][j]*CO_CL_WL_D[i][j];
        }
    }
}
Fisher_M[FX][FY] += sum;

Note that this code is memory bound, not computation expensive, so the perfomance gain by parallelization may be smaller than expected (and depends on your hardware).

Ps: Why do you need this condition if( CO_CL_WL[i][j] != 0 || CO_CL_WL_D[i][j] != 0) ? If any of them is zero, the sum will not change. If you remove it, the compiler can make much better vectorized code.

Ps2: In the schedule(dynamic, num_threads) clause the second parameter is the chunk size not the number of threads used. I suggest removing it in your your case. If you wish to specify the number of threads used, please add num_threads clause or use omp_set_num_threads function.